Apparatus, method, recording medium to extract unlabeled data for updating a model

An information processing method performed by at least one processor, the information processing method comprising:

• acquiring labeled data and unlabeled data, the labeled data being image data to which a ground-truth label and attribute information indicating an attribute of an image are attached, the unlabeled data being image data to which attribute information indicating an attribute of an image is attached;
• calculating a distance between the labeled data and the unlabeled data based on the attribute information of the labeled data and the attribute information of the unlabeled data;
• extracting, from multiple pieces of unlabeled data each being the unlabeled data, unlabeled data smaller in the distance from the labeled data than another unlabeled data;
• and
• updating a model parameter of a learning model based on the labeled data and the unlabeled data extracted, wherein
• the extracting of the unlabeled data comprises extracting the unlabeled data such that a maximum value of the distance of the unlabeled data extracted increases as training of the learning model progresses,
• the information processing method further comprises training the learning model in separate steps,
• the training of the learning model includes the updating of the model parameter, and
• the extracting of the unlabeled data comprises, in a first step, extracting one or more first pieces of the unlabeled data in which a first inter-attribute distance from the labeled data is closer than other pieces of the unlabeled data, and to, in a second step, further extracting one or more second pieces of the unlabeled data in which a second inter-attribute distance from the labeled data is greater than the first inter-attribute distance.

A computer-readable, non-transitory recording medium having recorded thereon a program for causing a computer to

• acquire labeled data and unlabeled data, the labeled data being image data to which a ground-truth label and attribute information indicating an attribute of an image are attached, the unlabeled data being image data to which attribute information indicating an attribute of an image is attached;
• calculate a distance between the labeled data and the unlabeled data based on the attribute information of the labeled data and the attribute information of the unlabeled data;
• extract, from multiple pieces of unlabeled data each being the unlabeled data, unlabeled data smaller in the distance from the labeled data than another unlabeled data;
• and
• update a model parameter of a learning model based on the labeled data and the unlabeled data extracted, wherein
• the program further causes the computer to extract the unlabeled data such that a maximum value of the distance of the unlabeled data extracted increases as training of the learning model progresses,
• the program further causes the computer to train the learning model in separate steps, and
• the program further causes the computer to, in a first step, extract one or more first pieces of the unlabeled data in which a first inter-attribute distance from the labeled data is closer than other pieces of the unlabeled data, and to, in a second step, further extract one or more second pieces of the unlabeled data in which a second inter-attribute distance from the labeled data is greater than the first inter-attribute distance.